Digest · Topology and routing · read after Chapter 5
導讀 · 拓樸與路由 · 第 5 章後閱讀
LEO Topology Design Under Real-World Deployment Constraints真實部署限制下的 LEO 拓樸設計
Stable links remain feasible for a complete orbit. Using real Starlink ephemerides, the study builds shortcuts that reduce propagation delay and hop count. The comparison also shows how partial deployment can split a three-link grid into separate components.穩定鏈路在整個軌道週期內持續可行。研究以真實 Starlink 星曆建立捷徑，減少傳播延遲與跳數，並顯示部分部署如何使三鏈路網格分成多個連通分量。
If you read nothing else
只讀這一段也可以
Classic ISL topology papers assume a complete, symmetric constellation with static links. The existing Starlink shell is partial, uneven and changes daily. The authors filter candidate links from actual ephemerides. They construct deterministic cyclic LSL and simulated-annealing SA topologies from stable candidates.
The baseline is a +Grid adapted to existing neighbors. With three terminals per satellite, the number Starlink publishes, a 3-ISL grid splits into separate connected components on the real ephemerides. Delay includes only satellite-to-satellite propagation. Ground links, node delay and scheduling lie outside the model.
The work is a preprint.
所有經典的 ISL 拓樸論文都假設完整、對稱、鏈路靜態的星座，但真正存在的 Starlink 殼層是部分部署、分布疏密相異，而且每天持續變化。他們從這組候選鏈路建出兩種拓樸，一是稱為 LSL 的確定性循環樣式，二是稱為 SA 的模擬退火搜尋，再與一個依現有鄰居調整過的 +Grid 比較。若每顆衛星只有三個終端，也就是 Starlink 公布的數字，3-ISL 網格在真實星曆上甚至分成多個連通分量。限制在於這裡的延遲是衛星之間的純傳播時間，計算範圍排除地面段、節點延遲與排程器，而且這是一篇預印本。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Orbital period | Ali and Zhang use 90 daily Space-Track TLE sets for Shell 1 from October to December 2024. SGP4 propagates each satellite through one period. Stable pairs remain below both the line-of-sight limit and range cap throughout that period. |
| Path length and hops | Across real-data satellite pairs, +Grid averages 89.0 ms and 23.4 hops. LSL averages 52.9 ms and 8.2 hops, while SA gives 49.3 ms and 12.0 hops. At 100 Gbps per ISL direction, LSL max-min-fair capacity over 100 flows is 17.5 Tbps. +Grid gives 7.5 Tbps. [34, §5.5] |
| 傳播延遲 | Ali 與 Zhang 取用 Shell 1 連續 90 天（2024 年 10 月至 12 月）每日一筆的 Space-Track TLE，以 SGP4 傳播，只保留「穩定鏈路」，也就是在一個軌道週期內最大距離始終低於視線極限與距離上限的衛星對。 |
| 路由效能 | 在真實資料上，+Grid 對所有衛星對的平均值是 89.0 ms 與 23.4 跳，LSL 達到 52.9 ms 與 8.2 跳，SA 達到 49.3 ms 與 12.0 跳。 |
| 容量比較 | 在每個 ISL 方向 100 Gbps 的設定下，LSL 對 100 條流的最大最小公平容量為 17.5 Tbps，+Grid 則為 7.5 Tbps[34, §5.5]。 |
1. Where this paper sits1. 這篇論文的位置
這篇預印本則是事實庫裡第一篇用「觀測到的星座」取代申報文件的拓樸研究。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Topology baseline | Chapter 5 presents the +Grid and the CoNEXT 2019 motifs as a design space calibrated to filings whose full deployment remained unrealized. This preprint is the first topology study in the fact base that replaces the filing with the constellation as observed. |
| Plane configuration | Motifs use 24 planes of 66 satellites and four ISLs per satellite. Hypatia, StarryNet and the emulators use the August 2019 layout of 72 by 22. |
| Constellation configuration | Ali and Zhang start from the same 72 by 22 provisioning for Shell 1. The layout has 1,584 slots. Section 1 reports partial deployment since the 2019 launches and a plan change in early 2022. Active satellites change daily through orbit raising, retirement and relocation. |
| Model configuration | They also take the terminal count from the Starlink website, three optical links per satellite, rather than the four that the whole modeling era used. This is why a 3-ISL evaluation sits beside the 4-ISL one. |
| Evaluation conditions | The paper is a preprint dated 8 February 2026 and awaits refereed status. Each result carries a preprint tag. Inputs are public ephemerides, and fact-base assembly verified the headline values against the text. |
| 評估條件 | 第 5 章把 +Grid 與 CoNEXT 2019 的 motif 當成一個設計空間來講，而這個設計空間是校準在完整部署仍待實現的申報文件上。 |
| 軌道面配置 | Bhattacherjee 與 Singla 的 motif 假設 24 個軌道面各 66 顆、每顆四條 ISL，而 Hypatia、StarryNet 與各模擬平台則凍結了 2019 年 8 月修正案的 72 面 × 22 顆布局。 |
| 星座配置 | Ali 與 Zhang 從同一份 Shell 1 的 72 × 22 配置出發，也就是 1,584 個席位，但他們在第 1 節指出，自 2019 年發射以來完整部署仍待實現，計畫在 2022 年初已經改變，而且每天的活躍衛星集合都隨著升軌、退役與調位而變動。 |
| 終端數假設 | 他們也採用 Starlink 網站上的終端數，即每顆衛星三條光學鏈路，對照整個建模時代慣用的四條，這就是為什麼 3-ISL 的評估與 4-ISL 的評估並列。 |
| 評估條件 | 這篇論文是 2026 年 2 月 8 日的預印本，尚同儕審查狀態待確認，因此下文每個數字都帶著這個標記，其輸入是公開星曆，而主要數字在整理事實庫時也已對照原文核對過。 |
2. The problem, measured2. 問題，用量測說話
The measurement here is of geometry rather than of traffic. The authors propagate every active Shell 1 satellite through one orbital period from daily Space-Track TLEs. They then compute pairwise distances. This exposes two facts that the synthetic constellations hide.
The first is that the shell is far from the regular lattice the +Grid assumes. Thus, the "same-index satellite in the adjacent plane" that +Grid wires to may occupy an empty slot on a given day. The adapted +Grid must fall back to the nearest available neighbor, stretching paths.
這裡量的是幾何，流量則位於量測範圍之外。作者利用 Space-Track 為每顆衛星每天發布的一筆 TLE 快照，把 Shell 1 所有活躍衛星傳播一個軌道週期並計算兩兩距離，由此暴露出合成星座所掩蓋的兩件事。第一，這個殼層與 +Grid 假設的規則格點有顯著差異，所以 +Grid 要連的「相鄰軌道面上同序號的衛星」在某一天可能對應一個空席位，調整過的 +Grid 只能退而連到最近的可用鄰居，路徑因而被拉長。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Orbital period | Pairs close at one instant can leave range within the period. Satellites in separate planes approach near plane crossings and separate near the equator. Section 3 describes this behavior, and Chapter 2 derives it from cos i scaling. |
| Evaluation conditions | For the provisioned 72 by 22 layout, the authors report maximum intra-orbit distance around 1,950 km and inter-orbit distance around 600 km. Chord calculations give 1,969 km and 604 km. The latter falls to 363 km at the highest latitude of a 53° shell. |
| Power budget | The 80 km atmospheric exclusion gives a 5,016 km line-of-sight ceiling, matching Chaudhry and Yanikomeroglu. Stable links remain below this ceiling and any tighter power-limited cap throughout the orbit. [17, Eq. 1] |
| 星座配置 | 第二，許多在某一刻很近的衛星對會在週期內漂出範圍，因為各異軌道面上的衛星在軌道面交會處靠近、在赤道附近分離，論文在第 3 節陳述了這點，第 2 章也從軌道面間距的 cos i 比例關係導出同樣的結論。 |
| 評估條件 | 對於配置的 72 × 22 幾何，作者說最大軌道內距離約 1,950 km、最大軌道間距離約 600 km，下方的計算表用弦長公式重算得到 1,969 km 與 604 km，而在 53° 殼層的最高緯度處會縮到 363 km。 |
| 穩定鏈路距離上限 | 扣除 80 km 大氣層後的視線上限與 Chaudhry 和 Yanikomeroglu 導出的 5,016 km 相同[17, Eq. 1]，所以一條穩定鏈路必須在整個軌道週期內都低於這個上限，並且低於一個更小的功率受限距離上限。 |
| Question問題 | Calculation算式 | Result結果 |
|---|---|---|
| Intra-plane neighbor distance, 22 satellites per plane at 550 km每面 22 顆、550 km 時的軌道內鄰居距離 | \(2R\sin(\pi/22),\ R = 6{,}921\ \text{km}\) | 1,969 kmcomputed. Matches the paper's "about 1,950 km" [34, §3]計算值。與論文「約 1,950 km」相符[34, §3] |
| Same-index adjacent-plane distance at the equator, 72 planes72 個軌道面時，赤道處相鄰面同序號衛星的距離 | \(2R\sin(2.5^\circ)\) | 604 kmcomputed. The paper's "about 600 km" [34, §3]計算值。論文「約 600 km」[34, §3] |
| Same pair at the highest latitude of a 53° shell同一對衛星在 53° 殼層最高緯度處的距離 | \(604\ \text{km}\times\cos 53^\circ\) | 363 kmcomputed. Ratio ≈ cos i from Chapter 2計算值。比例 ≈ cos i，見第 2 章 |
| One-way propagation over the long intra-plane hop沿長的軌道內跳段的單向傳播時間 | \(1{,}969\ \text{km}\times 3.336\ \mu\text{s/km}\) | 6.57 mscomputed. Vacuum propagation from Chapter 2計算值。真空傳播，見第 2 章 |
3. The design3. 設計
The design has three parts, a feasibility filter, two topology constructors and an update rule. Figure A shows the filter and the LSL pattern.設計分為三部分：一個可行性過濾器、兩個拓樸建構器，以及一條更新規則。作者把這當成保守的選擇，並把暫時可行的鏈路留作後續工作[34, §3]。圖 A 畫出過濾器與 LSL 樣式。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Constellation configuration | Equation 2 accepts a pair whose maximum orbital-period distance stays below the smaller of its range cap and Equation 1 line-of-sight distance. Only these stable links enter the graph. The authors call this a conservative choice and leave transient links for future work. [34, §3] |
| Evaluation conditions | LSL preserves each intra-orbit ring and adds cyclic inter-orbit links. Along the orbit, plane separation cycles from D through D minus 1 to 1. Thus, each satellite owns one long shortcut and the pattern mixes short local links with long ones. |
| Evaluation conditions | D is 9 for Starlink and 4 for Kuiper [34, Tab. 4]. |
| Delay comparison | SA searches the stable-link graph with 200,000 initial and 100,000 incremental iterations. A Dijkstra evaluation costs 2.09 s, implying about 58 hours for 100,000 iterations. SA therefore optimizes constant-time surrogates: normalized mean link length, long-link fraction and edge utilization. The paper checks their correlation with delay and hop objectives, leaving a proof open. [34, Tab. 1] |
| Constellation configuration | The update rule in Section 4.4 recomputes incrementally each day as satellites appear or vanish. The incremental LSL stays within roughly 1 to 5 ms of a full recomputation under shell growth and shrinkage [34, Figs. 11, 12]. |
| 星座配置 | 過濾器就是論文的 Eq. 2：兩顆衛星之間的候選鏈路若在一個週期內的最大距離至多為距離上限與 Eq. 1 視線距離兩者中的較小值，就算穩定，而且只有穩定鏈路才進入圖中。 |
| 軌道面配置 | 第一個建構器 LSL 是一種確定性的循環樣式，保留軌道內的環，再加入軌道間鏈路，其軌道面間隔沿著軌道依序循環 D、D−1、……、1，使每顆衛星擁有一條長捷徑，整個樣式混合了短的局部鏈路與長鏈路。 |
| 評估條件 | Starlink 的 D 為 9、Kuiper 為 4[34, Tab. 4]。 |
| 延遲比較 | 第二個建構器 SA 在穩定鏈路圖上執行模擬退火，初始 200,000 次、增量 100,000 次迭代，但由於對一個候選拓樸做一次 Dijkstra 評估要 2.09 s、每 100,000 次迭代約需 58 小時[34, Tab. 1]，它改為最佳化常數時間的替代指標，即正規化平均鏈路長度、長鏈路比例與邊的使用率，而這些指標與真實延遲、跳數目標的相關性，論文是用經驗檢驗的方式確認的。 |
| 星座配置 | 第 4.4 節的更新規則在衛星出現或消失時每天做增量重算，增量式 LSL 在殼層成長與縮減時與完整重算的差距大致在 1 到 5 ms 之內[34, Figs. 11, 12]。 |
Original figure note原始圖說出處
Drawn for this primer from Ali & Zhang, arXiv 2602.07756, §3, §4, Eqs. 1–2, Table 4.4. Evaluation setup4. 實驗設定
This trace-driven simulation evaluates geometric feasibility and propagation paths. Capacity is a max-min-fair allocation over 100 random satellite pairs. Laser acquisition time appears only as background, described as seconds to minutes depending on hardware and relative motion [34, §2], and is excluded from topology cost. This is defensible precisely because the stable-link filter means links are held for the whole orbit and only change at the daily update. The 15-second Starlink scheduler is outside the discussion. This is consistent with a satellite-to-satellite study that stays within satellite-to-satellite paths.這是一個由軌跡資料驅動的幾何模擬，評估範圍聚焦於幾何可行性與傳播路徑。容量是對 100 對隨機衛星的最大最小公平分配。雷射擷取時間只作為背景出現，描述為依硬體與相對運動而異的數秒到數分鐘[34, §2]，列於拓樸成本範圍之外。這一點站得住腳，正因為穩定鏈路過濾器意味著鏈路在整個軌道週期內持續保持，只在每日更新時才改變。Starlink 的 15 秒排程器位於論文範圍之外，對一個範圍限於衛星間路徑的衛星對衛星研究而言這是一致的。
| Measure or comparison參數或比較項目 | Values, conditions and evidence數值、條件與證據 |
|---|---|
| Evaluation conditions | Real ephemerides cover 90 days from October to December 2024, with one daily Space-Track TLE per Shell 1 satellite and SGP4 propagation. Comparison tables also use synthetic full 72 by 22 Starlink and synthetic Kuiper constellations. [34, §5.1] |
| Model configuration | Feasible links use constant full-duplex 100 Gbps per direction up to roughly 8,000 km. The main evaluation allows four ISLs per satellite. A Starlink-matched variant allows three. [34, §5.4.2] [34, §2, §5.2] |
| Latency comparison | Latency averages shortest-distance propagation paths over all satellite pairs. Ground stations, user terminals, per-hop node delay and queueing lie outside its scope. [34, §5.3] |
| 星座配置 | 星曆是真實的：Starlink Shell 1 從 2024 年 10 月到 12 月共 90 天、每顆衛星每天一筆的 Space-Track TLE，以 SGP4 傳播[34, §5.1]，而比較表格另外使用合成的完整 72 × 22 Starlink 與合成的 Kuiper。 |
| 模型配置 | ISL 模型極為簡單，任何可行鏈路在約 8,000 km 以內都固定為每方向 100 Gbps 全雙工[34, §5.4.2]，主要評估每顆衛星四條 ISL，Starlink 對應版本則為三條[34, §2, §5.2]。 |
| 延遲比較 | 延遲定義為所有衛星對之間沿最短距離路徑的平均傳播延遲[34, §5.3]，因此地面站、用戶終端、其他論文假設的 1 ms 或 10 ms 每跳節點延遲及排隊，皆位於模型範圍之外。 |
5. What was measured5. 量到了什麼
| Claim宣稱 | Setup and caveat設定與附註 |
|---|---|
| +Grid on real Shell 1 ephemerides averages 89.0 ms and 23.4 hops. LSL 52.9 ms (40% lower) and 8.2 hops (65% fewer). SA 49.3 ms (45%) and 12.0 hops (49%) [34, §5.5.1].在真實 Shell 1 星曆上 +Grid 平均 89.0 ms、23.4 跳。LSL 為 52.9 ms（低 40%）、8.2 跳（少 65%）。SA 為 49.3 ms（45%）、12.0 跳（49%）[34, §5.5.1]。 | 4 ISLs per satellite. Mean over all satellite pairs. Propagation only, ground segment excluded. The abstract's "up to 45%" is SA, the 52.9 ms figure is LSL.每顆衛星 4 條 ISL。對所有衛星對取平均。僅傳播延遲、地面段排除。摘要中的「最多 45%」指 SA，52.9 ms 指 LSL。 |
| Max-min-fair capacity over 100 pairs: LSL 17.5 Tbps against +Grid 7.5 Tbps, "nearly 2.3×". SA about 12.5 Tbps, "nearly 1.7×" [34, §5.5.2].對 100 對衛星的最大最小公平容量：LSL 17.5 Tbps、+Grid 7.5 Tbps，「接近 2.3 倍」。SA 約 12.5 Tbps，「接近 1.7 倍」[34, §5.5.2]。 | 100 Gbps per direction on every link regardless of range. The gain comes from fewer hops sharing each link, with laser rates held fixed.每條鏈路每方向都是 100 Gbps、固定於各種距離。增益來自每條鏈路被較少跳數分攤，雷射速率則維持固定。 |
| Synthetic full Starlink, 4 ISLs: LSL 46.9 ms / 8.8 hops. SA low-hop 61.2 ms / 8.5. +Grid 60.6 ms / 23.5. Motif 48.9 ms / 21.5. Theoretical floor 36.1 ms [34, Tab. 2].合成完整 Starlink、4 條 ISL：LSL 46.9 ms / 8.8 跳。SA 低跳數版 61.2 ms / 8.5。+Grid 60.6 ms / 23.5。Motif 48.9 ms / 21.5。理論下限 36.1 ms[34, Tab. 2]。 | Motif is evaluated only here because it requires full symmetric deployment and is inapplicable to the partial real shell.Motif 只在這裡評估，因為它需要完整對稱部署，其適用範圍限於完整對稱星座。 |
| With 3 ISLs on synthetic Starlink: 3-ISL grid 146.4 ms / 36.6 hops. LSL 80.0 ms / 13.3. SA 61.6 ms / 15.9 [34, Tab. 3]. On real data the 3-ISL grid is disconnected and excluded.合成 Starlink 上 3 條 ISL：3-ISL 網格 146.4 ms / 36.6 跳。LSL 80.0 ms / 13.3。SA 61.6 ms / 15.9[34, Tab. 3]。在真實資料上 3-ISL 網格分成多個連通分量而被排除。 | Three terminals is the Starlink website figure. With one ring link removed the grid loses the redundancy that keeps a partial shell connected.三個終端是 Starlink 網站上的數字。去掉一條環鏈路後，網格就失去讓部分部署殼層保持連通的冗餘。 |
| Daily link breakage: SA 1.0%, LSL 1.3%, +Grid 1.2% [34, Fig. 10]. Incremental LSL within about 1–5 ms of full recomputation [34, Figs. 11–12].每日鏈路斷裂比例：SA 1.0%、LSL 1.3%、+Grid 1.2%[34, Fig. 10]。增量式 LSL 與完整重算相差約 1–5 ms[34, Figs. 11–12]。 | Breakage means a stable link of day d loses stability on day d+1. The designs are "comparable". Thus, the delay gain keeps churn at a comparable level.斷裂指第 d 天的穩定鏈路在第 d+1 天失去穩定性。三種設計「相當」，所以延遲的增益仍維持相近的鏈路變動率。 |
6. Limits and open questions6. 限制與待解的問題
- Preprint status. The paper is arXiv 2602.07756v1 and awaits peer review. Thus, the fact base tags every figure as preprint even though the headline numbers were checked against the text.預印本狀態。本文是 arXiv 2602.07756v1，同儕審查狀態仍待確認，所以即使主要數字已對照原文核對，事實庫仍把每個數字標記為預印本。
- The 89.0 ms value measures one-way inter-satellite propagation. The metric averages one-way satellite-pair propagation, with zero node delay. Ground stations and bent-pipe legs lie outside its scope. Chapter 3 measures Starlink idle RTT around 40–50 ms for bent-pipe users and 100 ms for ISL-dependent islands. Each quantity requires its own path scope and delay definition.89.0 ms 量測的是衛星間單向傳播。它是對所有衛星對的單向平均，範圍排除地面站、彎管段及節點延遲。而第 3 章量到的 Starlink 閒置 RTT 下限，彎管用戶為 40 到 50 ms、依賴 ISL 的島嶼約 100 ms，比較時須保留各自的路徑範圍與延遲定義。
- Rate is independent of range. The model holds 100 Gbps per direction constant to about 8,000 km. Chapter 2 instead shows a 10 cm aperture losing 8 dB between 2,000 and 5,000 km. The 8,000 km range also exceeds the 5,016 km line-of-sight limit in Equation 1. Thus, the range cap rather than the rate assumption is doing the work.速率固定於各種距離。到約 8,000 km 都固定為每方向 100 Gbps，這與第 2 章的鏈路預算矛盾，同一個 10 cm 孔徑在 2,000 到 5,000 km 之間會損失 8 dB。而 8,000 km 本身也超過論文自己 Eq. 1 所強制的 5,016 km 視線距離，所以真正起作用的是距離上限，速率則固定於各種距離。
- Setup delay is sidestepped, outside the model. Links stay fixed within each full orbit. Daily incremental updates still change 1.0–1.3% of links. Each change costs the 2 to 30 s acquisition of Chapter 2 that the paper leaves outside its cost model.穩定鏈路避開週期內的建立成本。穩定鏈路在整個軌道週期內保持，去除了週期內的重組，但每日的增量更新仍會改變 1.0 到 1.3% 的鏈路，而每次改變都要付出第 2 章所說的 2 到 30 秒擷取時間，論文將其排除於成本之外。
- The 15-second schedule and the user link are absent. This is appropriate for a satellite-to-satellite study but means the design requires additional models to establish end-to-end behavior. A topology with 8.2 hops still ends in a user link that reconfigures every 15 s.15 秒排程與用戶鏈路位於模型範圍之外，對衛星對衛星的研究來說這是合理的，但這也意味著設計須加入額外模型才能評估端到端行為。一個 8.2 跳的拓樸最後仍接到一條每 15 秒重組一次的用戶鏈路。
- Baselines are uneven. Motif is evaluated only on synthetic 4-ISL constellations. Real-data 3-ISL grids disconnect and are excluded. With three ISLs, LSL performs worse than SA at 77–80 ms. Thus, the strongest real-data comparison is LSL and SA against an adapted +Grid only.各基準的評估範圍各異。Motif 只限於合成的 4-ISL 情境，3-ISL 網格因分成多個連通分量而在真實資料上被排除，而 LSL 自身在三條 ISL 下也比 SA 差（77 到 80 ms），所以最有力的真實資料比較只有 LSL 與 SA 對調整過的 +Grid。
- SA's surrogates are validated by correlation. Capacity uses only 100 pairs. The paper leaves error bars for SA optimality and the 2.3× capacity ratio open.SA 的替代指標只以相關性驗證，容量數字也只用 100 對衛星，所以 SA 的最佳性與 2.3 倍的容量比在論文中的誤差範圍皆仍待補充。
7. How it connects7. 與其他內容的關聯
- Section 5.2 compares +Grid, motifs and stable links on real TLEs. This paper supplies the 89.0 / 52.9 / 49.3 ms comparison.第 5.2 節：這篇論文是 +Grid、motif 與穩定鏈路三方比較中「真實 TLE 上的穩定鏈路」這一支，並提供 89.0 / 52.9 / 49.3 ms 三組數字。
- Section 5.7 uses the three-terminal case and disconnected 3-ISL grid to reassess legacy +Grid results.第 5.7 節：從四個終端變成三個以及分成多個連通分量的 3-ISL 網格，是該章主張舊的 +Grid 結果需要重新解讀的主要證據。
- Section 5.6 views the daily 1.0–1.3% link breakage from the topology side. The MobiCom 2023 digest examines related orbital operations.第 5.6 節：每日 1.0 到 1.3% 的鏈路斷裂，是從拓樸一側看到的軌道機動，MobiCom 2023 的導讀則從營運一側描述同一件事。
- Section 4.2 cites this paper for intra-plane and inter-plane distances: 1,969 km and 604–363 km. It also cites 100 Gbps per direction and 3 terminals per satellite.第 4.2 節：參數表中軌道內與軌道間 ISL 長度（1,969 km、604 到 363 km）、每雷射速率（每方向 100 Gbps）與每衛星終端數（3）的各列都引用這篇論文。
- CoNEXT 2019 motifs digest: the 54% motif gain was computed on a full 24 by 66 shell with four ISLs. This is exactly the assumption this paper shows to be inapplicable to the real partial shell.CoNEXT 2019 motif 導讀：54% 的 motif 增益是在完整的 24 × 66 殼層、四條 ISL 下算出的，而這正是本文顯示需要完整部署的假設，與真實部分部署殼層各有條件。
- Handley 2018 digest uses five ISLs at 1,150 km. TinyLEO digest uses three at 200 Gbps and records Starlink's proprietary topology. This paper sits between them as the only one built on observed positions.Handley 2018 導讀與TinyLEO 導讀：Handley 在 1,150 km 的申報上接了五條 ISL，TinyLEO 接三條 200 Gbps 的鏈路並聲明 Starlink 的拓樸仍待公開。本文介於兩者之間，是唯一建立在觀測位置上的研究。
Key numbers關鍵數字
| Number | 數字 | Meaning (setup) | 意義（設定） |
|---|---|---|---|
| 90 days | Daily Space-Track TLEs for Shell 1, Oct–Dec 2024, propagated with SGP4 [34, §5.1] | Shell 1 的每日 Space-Track TLE，2024 年 10–12 月，以 SGP4 傳播[34, §5.1] | |
| 5,016 km | Line-of-sight limit with an 80 km atmosphere, the ceiling in Eq. 1 | 扣除 80 km 大氣層的視線極限，Eq. 1 的上限 | |
| 1,969 / 604 / 363 km | Intra-plane, equatorial inter-plane and high-latitude inter-plane distances for 72×22 [computed] | 72×22 的軌道內、赤道處軌道間與高緯度軌道間距離 [計算值] | |
| 100 Gbps | Per direction per ISL, constant with range [34, §5.4.2] | 每條 ISL 每方向，固定於各種距離[34, §5.4.2] | |
| 4 → 3 | ISLs per satellite in the main and Starlink-matched evaluations [34, §2, §5.2] | 主要評估與 Starlink 對應評估中的每衛星 ISL 數[34, §2, §5.2] | |
| 89.0 ms / 23.4 | +Grid mean one-way delay and hops over all satellite pairs, real data [34, §5.5.1] | +Grid 在真實資料上對所有衛星對的平均單向延遲與跳數[34, §5.5.1] | |
| 52.9 ms / 8.2 | LSL on the same data, 40% and 65% lower [34, §5.5.1] | LSL 在同一資料上，各自低 40% 與 65%[34, §5.5.1] | |
| 49.3 ms / 12.0 | SA low-delay variant, 45% and 49% lower [34, §5.5.1] | SA 低延遲版本，各自低 45% 與 49%[34, §5.5.1] | |
| 17.5 vs 7.5 Tbps | Max-min-fair capacity over 100 pairs, LSL vs +Grid, "nearly 2.3×" [34, §5.5.2] | 對 100 對衛星的最大最小公平容量，LSL 對 +Grid，「接近 2.3 倍」[34, §5.5.2] | |
| 146.4 ms / 36.6 | 3-ISL grid on synthetic full Starlink [34, Tab. 3]. Disconnected on real data | 合成完整 Starlink 上的 3-ISL 網格[34, Tab. 3]。在真實資料上分成多個連通分量 | |
| 1.0–1.3% | Links that break from one day to the next for SA, +Grid and LSL [34, Fig. 10] | SA、+Grid 與 LSL 從一天到下一天斷裂的鏈路比例[34, Fig. 10] | |
| 2.09 s | One Dijkstra evaluation per SA step, about 58 h per 100,000 iterations, hence surrogates [34, Tab. 1] | SA 每步一次 Dijkstra 評估，每 100,000 次迭代約 58 小時，因此改用替代指標[34, Tab. 1] |